For-each node

For-each node card

The For-each node maps a language model over an array of items: every item becomes one model call (the item as the prompt, your instructions as the system prompt), and the results are collected back into an array of { item, output } entries — in the same order as the input.

Use it to apply the same analysis, classification or transformation to a list — for example, score every row of a generated table, translate every line of a report, or summarize each search result — without writing a loop by hand.

Handles

Handle	Purpose
`Items`	The array to iterate over (a table, list or JSON array)
`Results`	The collected `[{ item, output }]` array

Configuration

Model — the language model that runs once per item.
Instructions — the system prompt, applied to every item (supports {…} templates). Describe what to do with each item and what to return.
Items path — a dot-path into the run context pointing at the array to iterate over (e.g. outputs.<node-id>.rows). Leave empty to use the incoming payload when it is an array.

Where the items come from

The array is resolved in this order:

If an Items path is set, read that path from the run context.
Otherwise use the incoming payload, when it is an array.

Non-string items are serialized to JSON before being passed as the prompt, so a table row or a nested object becomes a readable text prompt automatically. If no items resolve, the run fails fast with a clear message.

Example — score every lead

A Table Generation node produces { sheets: [...] }. A For-each node reads the rows and scores each one:

items:   payload.rows            (or outputs.<table-node>.sheets[0].rows)
model:   openai/gpt-4.1-mini
instructions: "Rate each lead's quality from 1 (cold) to 5 (hot).
              Reply with just the number."

The Results handle emits:

[
  { "item": "{"name":"ACME"}", "output": "5" },
  { "item": "{"name":"Globex"}", "output": "3" }
]

Wire that into a Table Generation node or an Agent to turn the raw scores into a finished report.

Notes

One durable step — the items run in order, not in parallel, so re-runs are deterministic and each call is billed separately.
If one item’s call fails, the batch stops and reports that error.
The card shows the model and the items source at a glance.

Table Generation — a common source of the array.
Agent — the single-item equivalent of this node.
Router — branch once, rather than iterate.